Static and Dynamic Robustness in Emergency- Phase Communication Networks Sean M. Fitzhugh 1 and Carter T. Butts 1,2 1 Department of Sociology 2 Institute of Mathematical Behavioral Sciences University of California, Irvine MURI AHM May 25 th , 2010 This material is based on research supported by the Office of Naval Research under award N00014-08-1-1015 as well as NSF award CMS-0624257
Outline Introduction: Network robustness and disaster 1. response Methodology: How to measure network robustness 2. Results and analysis: static case 3. Dynamic robustness: Methods and results 4. Concluding remarks 5.
Network Robustness and Disaster Response Disaster response teams carry out complex tasks which require extensive training and planning Typically operate in a volatile, chaotic environment Perform tasks that require substantial coordination Medical response/triage Resource allocation Search and rescue Evacuation
Network Robustness and Disaster Response Certain types of network structure are conducive to performing activities related to disaster response Locally centralized patterns of communication help large groups of individuals carry out complex tasks (Bavelas 1973) To enhance efficiency, certain actors can function as “information hubs”: may serve to coordinate actions of others
Network Robustness and Disaster Response Hub-dominated structure of observed WTC Radio networks is potentially efficient, but this structure creates vulnerabilities
Network Robustness and Disaster Response What happens if we eliminate the yellow node’s ties?
Network Robustness and Disaster Response Note how many nodes have been isolated with the removal of just one individual.
Network Robustness and Disaster Response What if we “remove” four more hubs?
Network Robustness and Disaster Response Dozens more nodes have been isolated.
Network Robustness and Disaster Response This network’s information hubs are weak points
Network Robustness and Disaster Response Why is vulnerability problematic for these networks? Without effective information transmission, tasks may be carried out in an unstructured, counterproductive, or inefficient manner (Auf Der Heide 1989) Worse, some tasks may be overlooked altogether Studying robustness patterns of communication networks allows us to see who is important in holding the network together Actors with predetermined coordinative roles or emergent coordinators?
Data: World Trade Center Radio Seventeen radio communications networks from the World Trade Center disaster (Butts and Petrescu- Prahova, 2005) Fixed-channel radio communication: groups are independent (no cross-channel radio communication), so we can think of them as separate organizations Networks reconstructed from transcripts Transmission from actor i to actor j is coded as an (i,j) edge Actors generally treat communication as dyadic Individual conversations dominate communication
Data: World Trade Center Radio Specialist networks: daily occupational routine involves emergency response Lincoln Tunnel Police, Newark command, Newark Police, Newark CPD, New Jersey Statewide Police Emergency Network (NJ SPEN1), NJ SPEN2, WTC Police, Port Authority Trans-Hudson (PATH) Police Non-specialist networks: lack daily involvement in emergency response, but were in some way involved with WTC response PATH radio communications, Newark operations terminals, Newark maintenance, PATH control desk, WTC operations, WTC vertical transportation, Newark facility management, WTC maintenance electric
Data: World Trade Center Radio Each network has a number of actors in institutionalized coordinative roles (ICR) Their formal role is to coordinate the actions of others in the network Transcribed labels such as “command”, “desk”, “operator”, “dispatch( er )”, “manager”, “control”, “base” Manage a variety of roles in these networks: assisting searches for personnel, advising units on traffic/closures, coordinating equipment/EMT/personnel distribution, forwarding information Will ICRs operate in their formal, institutionalized roles or will others adopt those roles?
How to Measure Network Robustness Test the robustness of a network by subjecting it to various “attacks” ( not literal attacks) Remove nodes from the network and see how well it holds up Two basic sequences of node failure: random and degree-targeted I also selectively target ICRs to assess their role in holding the network together (leads me to use four total variations of sequential node failure) Remove nodes until none remain in the network
How to Measure Network Robustness Random failure: remove nodes at random
How to Measure Network Robustness Degree-targeted failure: remove nodes in sequential order according to degree
How to Measure Network Robustness Random failure targeting ICRs: remove ICRs at random, followed by random removal of remaining nodes
How to Measure Network Robustness Degree-targeted failure targeting ICRs: remove ICRs in sequential order according to degree, followed by sequential removal of remaining nodes
How to Measure Network Robustness Connectivity: Who can reach whom? Isolate formation: Whose removal isolates others?
How to Measure Network Robustness Connectivity: Who can reach whom? Isolate formation: Whose removal isolates others?
How to Measure Network Robustness Connectivity: Who can reach whom? Isolate formation: Whose removal isolates others?
How to Measure Network Robustness Connectivity: Who can reach whom? Isolate formation: Whose removal isolates others?
Building Robustness Profiles We need a way to measure connectivity as a network progressively degrades Robustness scores: measure of a network’s declining connectivity as more and more of its nodes are removed Use simulation of node failure to obtain robustness scores After up many iterations, simulation yields expected mean connectivity as nodes are removed Let’s look at some examples for clarification…
Building Robustness Profiles Using either of the previous measures, plot the robustness curve to monitor network connectivity as more nodes fail
Building Robustness Profiles Use multiple plots to compare robustness of different series of node failures The area between curves tells us how network robustness differs across attacks
Building Robustness Profiles Take the integral of the curve to obtain a robustness score Connectivity Random failure: 0.4287 Random failure of ICRs: 0.0397
Building Robustness Profiles Robust example: Connectivity Random failure: 0.4159 Random failure of ICRs: 0.3579
Hypotheses With an understanding of how to measure network robustness, we can test some hypotheses Hypothesis 1: Specialist and non-specialist networks will be more robust to random failure than to random failure of ICRs Those with institutionalized roles will maintain those roles during the disaster response Hypothesis 2: Specialist networks will be less robust to loss of ICRs than non-specialist networks Trained for these types of tasks, specialists can consolidate their coordination needs onto a smaller number of people
Hypotheses Hypothesis 3: Degree targeted failure and degree- targeted failure of ICRs will produce similar robustness scores among specialist and non-specialist networks If ICRs occupy positions with the most ties, there should be no difference between the two attacks
Comparing Robustness Profiles Calculate robustness scores for all varieties of attacks (random, degree-targeted, and ICR-targeted) across measures of connectivity and isolate formation Use t-tests to compare scores across different dimensions (ICR vs. non-ICR failures, specialist vs non- specialist networks)
Static Robustness: Results Static robustness examines the time-aggregated networks Series of time-ordered communication events collapsed into a single network
Static Robustness: Results Hypothesis 1: Specialist and non-specialist networks will be more robust to random failure than to random failure of ICRs Hypothesis 2: Specialist networks will be less robust to loss of ICRs than non-specialist networks Specialist networks are significantly more robust to random failure than to random failure of ICRs t=4.2877, p=.0026 Among non-specialist networks, ICRs prove less crucial to preserving connectivity t=1.9004, p=.0991
Static Robustness: Results Hypothesis 3: Degree targeted failure and degree- targeted failure targeting ICRs will produce similar robustness scores among specialist and non-specialist networks Degree-targeted failure is significantly more damaging than degree-targeted failure of ICRs in specialist networks t=-2.4815, p=.0380 The difference between the two attacks is significant in non-specialist networks t=-4.0548, p=.0048
Dynamic Robustness: Methodology Ordinal nature of transcripts allows us to explore dynamic robustness Using the time-ordered sequence of communication to measure forward connectedness How would network unfold if certain actors were never present in the network?
Recommend
More recommend
